gcnii

Simple and Deep Graph Convolutional Networks (GCNII)

• Paper link: https://arxiv.org/abs/2007.02133
• Author's code repo: https://github.com/chennnM/GCNII.

Note that our implementation is little different with the author's in the optimizer. The author applied different weight decay coefficient on learnable paramenters, while TensorLayerX has not support this feature.

Dataset Statics

Dataset	# Nodes	# Edges	# Classes
Cora	2,708	10,556	7
Citeseer	3,327	9,228	6
Pubmed	19,717	88,651	3
Refer to Planetoid.

Results

TL_BACKEND="tensorflow" python gcnii_trainer.py --dataset cora --lr 0.01 --num_layers 64 --alpha 0.1 --hidden_dim 64 --lambd 0.5 --drop_rate 0.3 --l2_coef 0.001
TL_BACKEND="tensorflow" python gcnii_trainer.py --dataset citeseer --lr 0.01 --num_layers 32 --alpha 0.1 --hidden_dim 256 --lambd 0.5 --drop_rate 0.3 --l2_coef 0.001
TL_BACKEND="tensorflow" python gcnii_trainer.py --dataset pubmed --lr 0.01 --num_layers 16 --alpha 0.1 --hidden_dim 256 --lambd 0.4 --drop_rate 0.3 --l2_coef 0.001
TL_BACKEND="paddle" python gcnii_trainer.py --dataset cora --lr 0.01 --num_layers 64 --alpha 0.1 --hidden_dim 64 --lambd 0.5 --drop_rate 0.3 --l2_coef 0.001
TL_BACKEND="paddle" python gcnii_trainer.py --dataset citeseer --lr 0.01 --num_layers 32 --alpha 0.1 --hidden_dim 256 --lambd 0.4 --drop_rate 0.4 --l2_coef 0.001
TL_BACKEND="paddle" python gcnii_trainer.py --dataset pubmed --lr 0.01 --num_layers 16 --alpha 0.1 --hidden_dim 256 --lambd 0.5 --drop_rate 0.7 --l2_coef 0.001
TL_BACKEND="torch" python gcnii_trainer.py --dataset cora --lr --lr 0.01 --num_layers 64 --alpha 0.1 --hidden_dim 64 --lambd 0.5 --drop_rate 0.3 --l2_coef 0.001
TL_BACKEND="torch" python gcnii_trainer.py --dataset citeseer --lr 0.01 --num_layers 64 --alpha 0.1 --hidden_dim 64 --lambd 0.6 --drop_rate 0.4 --l2_coef 0.001
TL_BACKEND="torch" python gcnii_trainer.py --dataset pubmed --lr 0.01 --num_layers 64 --alpha 0.1 --hidden_dim 64 --lambd 0.4 --drop_rate 0.6 --l2_coef 0.001

Dataset	Paper	Our(pd)	Our(tf)	Our(tf)
cora	85.5	83.12(±0.47)	83.23(±0.76)	83.1(±0.9)
pubmed	73.4	72.04(±0.91)	71.9(±0.7)	71.4(±0.6)
citeseer	80.3	80.36(±0.65)	80.1(±0.5)	80.5(±0.3)

Notice that we do not use the same regularization method as the paper do, as TensorlayerX currently do not support it.